Practical notes from the engineers who build agents, copilots and AI products.
Reduce hallucinations with grounding, structured outputs, evals, confidence thresholds and human review where the stakes require it. No single technique removes them; a production system layers five defences so a fabricated answer is rare, detected when it happens, and never reaches a customer where it matters.
Safe unattended agents need scoped tools, spend and rate limits, approval gates for sensitive actions, evaluation suites and a full audit trail. None of these live in the prompt. They are enforced in code around the model, so a wrong decision is bounded, visible and reversible before it reaches a customer or a ledger.
Pin model versions, keep the eval suite ready and migrate behind routing with a regression run before switching traffic. A model deprecation plan turns a retirement notice into a scheduled change: evaluate the candidate, fix regressions, shift traffic gradually and keep the old version as fallback until shutdown.
GST compliance means e-invoice generation via the IRP, e-way bills and returns data, built into the finance module. The integration is a small amount of API work and a large amount of data discipline: every invoice must carry the fields the schema demands, and the IRN and QR code must be stored, printed and reconciled.
GraphRAG builds a graph of entities and relationships so questions about how things connect are answered better than by similarity search. It costs more to build and maintain than a vector index, so it earns its place only in relationship-heavy domains such as contracts, supply chains, org structures and clinical data.
Size from measured throughput per GPU on your workload, not from parameter counts; quantisation often halves the hardware needed. Pick the model that passes your evals, measure tokens per second at real prompt lengths, then divide peak demand by that number. Most deployments need less hardware than the quote suggests.
Collect one to three hundred real questions with verified answers; it becomes the accuracy benchmark and the regression suite. Mine questions people actually ask, verify each answer with an analyst, tag by type, score on results rather than query text, and rerun on every prompt, model or schema change.
Production needs evals, monitoring, cost controls, permissions, integration, documentation and an owner; tick all before launch. A proof of concept shows the model clears the bar on a sample; production means it keeps clearing it on live traffic, at a known cost, with someone accountable when it stops.
Fraud models score in real time, are validated on hold-out periods and monitored for drift as fraud patterns change. The model sits behind rules and in front of analysts: rules catch the known, the model catches the unusual, analysts label what they review, and those labels retrain the model on a schedule.